InDeCa

Calcium imaging enables monitoring of neuronal activity, but accurate deconvolution of calcium signals remains a major challenge. Conventional methods often yield noisy, continuous signals prone to overfitting, necessitating manual parameter tuning. Furthermore, post-hoc thresholding are required to infer spikes from deconvolved signals. We introduce InDeCa (Interpretable Deconvolution for Calcium Imaging), a novel algorithm that directly estimates binary spike trains while avoiding overfitting without the need for explicit regularization. This is achieved by convexifying the binary deconvolution problem with a thresholding process that naturally discourages overfitting. InDeCa also supports integer spike inference via temporal upsampling, and further reduces noise sensitivity through an error-weighting scheme. Additionally, InDeCa leverages the inferred spikes to iteratively refine and denoise a bi-exponential calcium kernel, yielding interpretable time constants that can be easily evaluated. By integrating thresholding, upsampling, error weighting, and spike-informed kernel estimation, InDeCa enhances both the accuracy and interpretability of spike inference. We validated InDeCa on simulated datasets and benchmarked it against popular deconvolution approaches using ground truth recordings spanning diverse calcium indicators and brain regions. Our results show that InDeCa consistently performs well across datasets with minimal parameter tuning, providing a robust and interpretable framework for spike inference in calcium imaging.